Cross-modal group-relation optimization for visible-infrared person re-identification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Pattern Recognition

Background:

• Visible-infrared person re-identification (VIPR) is crucial for intelligent transportation systems.
• Modal discrepancies between visible and infrared images hinder accurate person appearance discrimination.
• Existing methods often require complex networks to disentangle modal and appearance differences.

Purpose of the Study:

• To propose a novel method for optimizing modal discrepancies in VIPR without complex disentanglement.
• To introduce a cross-modal group-relation (CMGR) metric for measuring modal discrepancies.
• To develop a group-relation correlation (GRC) loss function to optimize CMGR.

Main Methods:

• Developed a cross-modal group-relation (CMGR) metric to capture relationships between individuals across visible and infrared modalities.
• Designed a group-relation correlation (GRC) loss function, utilizing Pearson correlations, to optimize CMGR.
• Integrated CMGR as a training-phase constraint to minimize modal discrepancies.

Main Results:

• The CMGR method demonstrated superior performance compared to state-of-the-art approaches on RegDB and SYSU-MM01 datasets.
• Achieved a notable improvement of over 7% in rank-1 identification rate on the RegDB dataset when using CMGR.
• The CMGR model effectively minimizes modal discrepancies without requiring execution during inference.

Conclusions:

• The proposed CMGR method offers an effective and efficient solution for addressing modal discrepancies in VIPR.
• Optimizing cross-modal group relations provides a promising direction for enhancing person re-identification accuracy.
• CMGR serves as a valuable training constraint, improving VIPR system performance significantly.